J Appl Physiol 120: 297–309, 2016.
First published November 5, 2015; doi:10.1152/japplphysiol.01110.2014.
Synthesis Review
Multilevel functional genomics data integration as a tool for understanding
physiology: a network biology perspective
Peter K. Davidsen,1 Nil Turan,2 Stuart Egginton,3 and Francesco Falciani1
1
Institute of Integrative Biology, University of Liverpool, Liverpool, United Kingdom; 2School of Biosciences, University of
Birmingham, Birmingham, United Kingdom; and 3School of Biomedical Sciences, Faculty of Biological Sciences, University
of Leeds, Leeds, United Kingdom
Submitted 17 December 2014; accepted in final form 4 November 2015
systems biology; data integration; genomics
MODELING IN PHYSIOLOGICAL SCIENCES
PHYSIOLOGY HAS EVOLVED AS a series of subdisciplines attempting to understand organismal function as a combination of
interacting components and systems. The last decade or so has
witnessed the development of systems biology as an investigative approach and its application in different areas of biology, ranging from engineering/synthetic biology (e.g., design
of bacterial strains with improved properties) to health sciences
(e.g., disease biomarker identification). Despite the lack of a
concise definition acceptable to the majority of the community
(30, 32), systems biology is frequently understood to be the
study of complex regulatory interactions in biological systems
using a holistic approach. This is often achieved by integrating
different experimental approaches within the conceptual
framework of a computational model (i.e. a mathematical
representation of a system that allows simulation of its behavior). Physiology is probably one of the few research areas in
biological sciences that has traditionally adopted such an ap-
Address for reprint requests and other correspondence: F. Falciani,
Centre for Computational Biology and Modelling, Institute for Integrative
Biology, Univ. of Liverpool, Crown St., Liverpool L69 7ZB, UK (e-mail:
[email protected]).
http://www.jappl.org
proach. It has long sought to understand the behavior of
complex biological processes and cellular systems using an
integrative approach and has extensively adopted mathematical
modeling in its tool set. Classical examples include August
Krogh’s tissue cylinder model of oxygen transport to skeletal
muscle (34j), and Huxley’s two-state cross-bridge model of
muscle contraction (26), which are still used by investigators
today. Indeed, this shows that using modeling to study a system
as a whole has been a key component of physiology from its
early days.
As often happens when a distinct discipline branches out of
another, there developed over time a separation of ideas based
in part on confusion arising from use of esoteric terminology,
similar concepts masked by unfamiliar language. There is,
therefore, a need for an overview of this relatively new discipline, to both emphasize the essential links with basic physiological principles and demystify the approach such that the
available tools may become more widely adopted in physiological research. The overall aim of this opinion-based review
is to describe, using concepts that will be intuitive to physiology researchers, different key methodologies available from
the systems biology community. In addition, we provide a
practical step-by-step guide for integrating multilevel data
8750-7587/16 Copyright © 2016 the American Physiological Society
297
Downloaded from http://jap.physiology.org/ by 10.220.33.1 on June 15, 2017
Davidsen PK, Turan N, Egginton S, Falciani F. Multilevel functional genomics
data integration as a tool for understanding physiology: a network biology perspective.
J Appl Physiol 120: 297–309, 2016. First published November 5, 2015;
doi:10.1152/japplphysiol.01110.2014.—The overall aim of physiological research
is to understand how living systems function in an integrative manner. Consequently, the discipline of physiology has since its infancy attempted to link multiple
levels of biological organization. Increasingly this has involved mathematical and
computational approaches, typically to model a small number of components
spanning several levels of biological organization. With the advent of “omics”
technologies, which can characterize the molecular state of a cell or tissue (intended
as the level of expression and/or activity of its molecular components), the number
of molecular components we can quantify has increased exponentially. Paradoxically, the unprecedented amount of experimental data has made it more difficult to
derive conceptual models underlying essential mechanisms regulating mammalian
physiology. We present an overview of state-of-the-art methods currently used to
identifying biological networks underlying genomewide responses. These are based
on a data-driven approach that relies on advanced computational methods designed
to “learn” biology from observational data. In this review, we illustrate an
application of these computational methodologies using a case study integrating an
in vivo model representing the transcriptional state of hypoxic skeletal muscle with
a clinical study representing muscle wasting in chronic obstructive pulmonary
disease patients. The broader application of these approaches to modeling multiple
levels of biological data in the context of modern physiology is discussed.
Synthesis Review
298
Multilevel Network Integration to Understand Physiology
within an analysis pipeline based around inferred interactions
of variables, modeled as a network based on statistical correlations, using a worked example in the field of physiological
sciences.
THE ADVENT OF FUNCTIONAL GENOMICS: A CHALLENGE
FOR PHYSIOLOGICAL MODELING
Davidsen PK et al.
of regulation. Perhaps surprisingly, the exponential growth in
publicly available omics data (35, 37) has not resulted in a
paradigm shift in our understanding of biology. The main
reason is the continuing challenge of integrating multivariate
datasets spanning multiple organization levels in a way that
allows the identification of discrete, small biomolecular networks that are truly important in the context of a specific
biological response (47). Such a task cannot be achieved
simply using unaided human interpretation. Rather, complex
computational techniques are needed that are able to integrate
and automatically “learn” the structure of a biological system.
Such a modeling framework is very different from what physiological sciences have traditionally employed.
TOWARD DATA-DRIVEN PREDICTIVE BIOLOGY
Although the modeling approach traditionally used by physiologists has been extremely successful, it suffers from severe
limitations when challenged with extensive omics data. For
example, physiological modeling relies to various degrees on a
mechanistic understanding of the biological system of interest
(16), which automatically limits the number of components
that can be included due to gaps in our current knowledge (19,
47). Moreover, estimation of model parameters, which is
usually a challenging task because of experimental limitations
(e.g., due to limited amount and quality of data), makes the
approach difficult to scale up to a larger number of components
and their interactions. Perhaps the most comprehensive example to date is modeling the cardiac cycle based on ion channel
kinetics (44).
With such large multivariate datasets, and little knowledge
about the way biomolecules are connected with each other and
to key phenotypic switches, the fundamental question is
whether or not we can “learn” the structure of biological
interaction networks from high-throughput data. Clearly, there
is a need for sophisticated computational tools that are able to
1) integrate genomewide measurements spanning multiple levels of biological organization (ranging from subcellular to
organ level); 2) identify key biomolecular components of the
system; and finally 3) statistically infer the way that these
biomolecules interact in a pairwise manner to generate an
observed biological response.
Central to these approaches is the concept of interaction
networks, a mathematical representation of a system of
biomolecules. Networks are commonly used to describe biological systems at different levels of complexity (e.g., metabolic and signal transduction networks). They can be descriptive models built using a wide spectrum of qualitative data
(e.g., biological knowledge of protein-protein interactions,
transcription factor binding, etc.), or they can be inferred from
quantitative measurements using complex computational models. In this case, they can be used to predict the behavior of the
system when perturbed.
In the following section, we summarize specific methodologies that can be applied to achieve such tasks.
COMPUTATIONAL APPROACHES FOR THE ANALYSIS OF
COMPLEX DATASETS
The process of modeling a biological system from complex
multilevel datasets can, for the sake of convenience, be divided
J Appl Physiol • doi:10.1152/japplphysiol.01110.2014 • www.jappl.org
Downloaded from http://jap.physiology.org/ by 10.220.33.1 on June 15, 2017
It is now clear that much of the complex mammalian
physiology or pathophysiology cannot be understood in sufficient detail through a reductionist approach alone. Although
this approach has proved valuable in explaining broad phenomena and individual mechanisms, linking multiple mechanisms and effects has proved challenging. For example, a
disease phenotype is rarely caused by a single dysfunctional
gene or protein. Instead, genetic variability, epigenetic modifications, posttranscriptional regulation mechanisms, etc., all
act in concert to determine a specific high-level phenotypic
response (43). The potential for such complex interaction
makes data interpretation much more complicated than originally envisioned, highlighting the need to move away from the
widespread “candidate gene” approach (39).
Triggered by the advent of genome sequencing, inspired by
the Human Genome Project, dramatic technological advances
within the last decade or so have led to increased throughput in
genomewide molecular analyses (i.e., genomics, epigenomics,
transcriptomics, proteomics, metabolomics). The comprehensive data acquisition tools developed to cope with large datasets
have allowed investigators to determine the molecular state of
cells, tissues, or even entire organs in a single experiment. Such
cost-effective omics approaches are now becoming prevalent in
biological and medical research and, consequently, have been
responsible for the generation of an incredibly large amount of
multivariate molecular data. A large proportion of this data is
available in the public domain via different online databases [e.g.,
NCBI Gene Expression Omnibus (5), EBI ArrayExpress (7), and
PRIDE (29)].
For example, mRNA microarray technology and more recently mRNA sequencing has provided insight into the transcriptional response of skeletal muscle to prolonged endurance
exercise training, highlighting a pronounced interindividual
variation at the molecular level that is consistent with the
heterogeneous response observed in a population of individuals
at the physiology level (31, 59). Statistical models built to
explain such variation as a function of gene expression data can
be exploited to identify underlying mechanisms controlling
tissue homeostasis. The transcriptional signatures identified in
such studies likely explain, at least in part, why some people
show great improvements in aerobic capacity [maximum O2
uptake (V̇O2max)], whereas others only experience smaller benefits, despite completing the same supervised exercise training
program. Another example of applying omics technology to
better understand human physiology concerns the quantification of individual levels of different proteins in health and
disease; by use of proteomics methodology, Holloway et al.
(24) were the first to investigate adaptations in human muscle
protein content to long-term exercise training on a large scale.
While such omics-based studies hint at the potential of a
data-driven approach, they also illustrate the difficulty in deriving conceptual models underlying the essential mechanisms
regulating physiology, as most are restricted to only one aspect
•
Synthesis Review
Multilevel Network Integration to Understand Physiology
Davidsen PK et al.
299
modeling principles, including ordinary differential equationbased methods (3), probabilistic modeling techniques (e.g.,
Bayesian theory models) (42, 64), state-space representation
models (23), and correlation-based methods. Note, while the
first three approaches are able to infer directed networks, their
capability is currently limited to inferring smaller networks
with few variables due to increased computational complexity
than possible with correlation approaches.
Importantly, this network inference part may potentially
benefit from a biomarker discovery phase, since it has been
shown that identified predictive variables are more likely to be
directly controlling important physiological processes and,
therefore, are good candidates to include in a network (47).
Similarly, whole networks can be used as an input for biomarker discovery procedures. It has been shown that often the
overall “activity” of a biological network (e.g., a specific
signaling pathway) is a better predictor than a few key individual genes, proteins, and/or metabolites. This implies that, in
the coming years, predictive biomarkers are more likely to
consist of a relatively large panel of measurements, possibly
spanning multiple levels of complexity within a pathway.
Current omics platforms are experiencing a rapid development,
as well as drop in costs, making routine collection of large
datasets a feasible option. Once a robust biological network has
been inferred, this may serve as a good basis for developing a
more conventional modeling approach to provide explanations
for observed phenomena that requires a mechanistic understanding of the system (Fig. 1C).
Finally, multiple computational models that initially were
developed independently can be integrated into larger and
more complex models, which allow responses to physiological/
pathological challenges to be simulated, thus integrating effects across multiple organs and/or pathways. These complex
Fig. 1. Schematic representation of the process involved in modeling a biological system by integrating knowledge from various sources, and complex
multilevel datasets. The process can be conceptually
subdivided into four distinct yet interconnected approaches (A–D). The experimental data used can
either be novel multivariate data generated in your
own (wet) laboratory, or taken from a public repository. These may then be used to identify predictive
biomarkers, i.e., variables that are predictive of a
defined outcome (e.g., response to exercise training), and also to inform development of important
networks that infer such outcomes; experimental
data and other source of biological knowledge may
also be useful in refining these representations of
complex interactions. Such networks may in turn
aid biomarker discovery, but are an essential precursor to computational models that are able to
explore underlying molecular mechanisms; again,
knowledge of specific biological issues may help in
their refinement. Finally, incorporation of these
models into larger scale analyses offer the potential
for in silico experimentation, whereby, e.g., the
effect of different therapeutic interventions on disease outcome may be tested.
J Appl Physiol • doi:10.1152/japplphysiol.01110.2014 • www.jappl.org
Downloaded from http://jap.physiology.org/ by 10.220.33.1 on June 15, 2017
into four conceptually distinct yet interconnected approaches
(Fig. 1).
The first approach is biomarker discovery (Fig. 1A), which
perhaps is most widely used in the analysis of functional
genomics datasets. Here the objective is to identify measurable
variables that are predictive of a given outcome (e.g., the
response to physical training in a population of individuals).
Such measurements can be molecular (e.g., gene expression,
protein levels, metabolite concentrations, genetic mutations)
and/or more traditional physiological endpoints (e.g., endurance, V̇O2max). The identification of predictive biomarkers can
be achieved by use of univariate and multivariate variable
selection strategies that aim to identify the most relevant
explanatory measurement(s), while developing a computational model that can accurately predict an outcome (60).
Univariate methods will test every variable (e.g., expression of
a given gene) on its own, whereas multivariate methods test
combinations of variables for their ability to explain a given
outcome. Clearly, multivariate approaches better resemble the
complex nature of biological networks and, therefore, are more
likely to provide insights into the mechanisms underlying a
complex phenotypic trait. Consistent with this notion, multigene biomarkers are often required for robust predictions in
independent datasets.
The second approach (Fig. 1B) consists of “reverse engineering” biomolecular networks from observational data (i.e.,
infer regulatory interactions between quantified biomolecules
based on mathematical principles). Here the overall aim is to
reconstruct the underlying structure of interactions between
biological molecules profiled using omics tools (ideally from
multiple data sources) and rigorous statistics. Such a network
inference framework can be achieved by applying a multitude
of approaches with varying underlying data assumptions and
•
Synthesis Review
300
Multilevel Network Integration to Understand Physiology
Inference of Biological Networks from Observational Data
Reverse engineering is an evolving field within networkbased systems biology. The rapid accumulation of omics data
in the postgenomic era has made it possible to infer (a.k.a.,
“reverse engineer”) models of cellular systems with the overall
aim of deducing the regulatory structure at a subcellular level.
Most of the network-based approaches that have been developed are in fact general and can be applied to any type of
experimental data. However, because the mRNA expression
profiling technology is the most mature omics discipline, most
applications have been developed to reconstruct transcriptional
networks (i.e., decode the mechanisms of transcriptional control). However, recently it has become apparent that, irrespective of the methodology used to generate data, to be able to
recapitulate the complex behavior of a biological system, it is
essential to integrate multiple types and scales of experimental
data (e.g., transcriptomic, proteomic, metabolomic).
Static vs. Dynamic Networks
Biological networks can be reconstructed from two different
types of experimental studies: either cross-sectional, e.g., representing a population of individuals at a given time (i.e.,
steady-state measurements following an experimental perturbation), or prospective, where the experimental data are available across a defined time course. In reverse engineering,
statistical inference of biological causality is an important goal
(10a). A simple example of causality could, for example, be a
transcription factor regulating the expression of several target
genes. Since determining cause and effects implies a direction
(i.e., the cause precedes the effect), inference of causality from
cross-sectional studies presents a challenge due to their static
nature, one that is less difficult when a time course is available.
Davidsen PK et al.
However, it must be stressed that both approaches are often
used in combination to, for example, integrate clinical crosssectional studies (thereby providing the researcher with a static
network representation) and experimental intervention studies
that can provide dynamic (prospective) models of the process
being studied. At present, most of the developed techniques
infer regulatory networks without any causality information
(likely due to the scarcity of time course datasets due to their
higher costs). However, a small number of causality detection
techniques have been proposed in the literature, such as dynamic Bayesian networks (48) and Granger causality (46). It is
also important to point out that true time course datasets can
only be developed when the sequence of events is measured
within the same cells/tissues. This is, for example, achieved
with imaging techniques that require complex molecular
probes and can typically be only applied to measure a relatively
small number of system components (14). Omics technologies
unfortunately are disruptive, so time course data derived using
these approaches are in fact a sequence of independent snapshots, which clearly limits the potential use of dynamic modeling tools.
A Primer for Network Inference Methods
The simplest method for inferring statistical relationships
between experimental variables is computing the pairwise
correlation coefficient across a large collection of heterogeneous samples (8). Usually such an approach is not able to
identify complex nonlinear dependencies and does not discriminate between direct and indirect connections. More complex
methods, such as the mutual information (MI) based ARACNE
(Algorithm for the Reconstruction of Accurate Cellular Networks) (38), also aim at establishing a statistical relationship
between pairs of variables but have a stronger theoretical
foundation. Because of the added mathematical complexity,
they can capture a broader range of biologically relevant
dependencies between variables, including nonlinear, nonmonotonic relationships; importantly, they can distinguish between direct and indirect relationships. ARACNE is a free tool
for which a Java-based graphical user interface exists; hence
investigators do not need any programming skills to use the
software.
ARACNE relies on estimating the probability that a variable
(e.g., the expression of a gene or a protein) assumes a certain
“state” (i.e., abundance), given the state of another biomolecule
(conditional probability). A number of alternative MI-based
implementations have been proposed during the last decade
[e.g., context likelihood relatedness (13), minimum redundancy/maximum relevance networks (41)], which mainly differ by the way inferred indirect relationships (so-called
“edges”) are removed once the dependencies between all pairs
of variables have been mathematically formulated. In such
analyses, unwanted indirect interactions occur by default if
there is strong correlation between biomolecule 1 and biomolecule 2, and between biomolecule 1 and biomolecule 3 in a
three-node clique (i.e., a triplet of connected variables).
An MI value of zero means that there is no dependency (i.e.,
no information flow) between two variables, whereas an MI
value of 1 indicates a perfect association between them, and,
therefore, a likely strong regulatory interaction between them.
For each inferred dependency, a P value is calculated based on
J Appl Physiol • doi:10.1152/japplphysiol.01110.2014 • www.jappl.org
Downloaded from http://jap.physiology.org/ by 10.220.33.1 on June 15, 2017
models are often referred to as decision support systems
because of their potential to provide information about the
expected outcome of a therapeutic intervention (Fig. 1D).
Several large international projects aiming at the development of such technology into systems medicine integrated
frameworks have been established so far, e.g., the Virtual
Physiological Human project funded by the European Commission 7th Framework Programme, which aims to aid clinically relevant research by establishing a framework for handling and integrating various mechanistic models spanning
different levels of organizational complexity (ranging from
molecular components to organ function). By unifying the
modeling languages employed across the different mathematical models included, parameters of a particular model in the
hierarchy can be processed by other appropriate models at a
lower hierarchical level. These global initiatives should be
considered long-term goals, aiming at understanding human
physiology quantitatively as a dynamic system.
Developing a comprehensive model of a biological system
requires integrating mechanistic and probabilistic inferences.
The mathematics for performing such a task is in its infancy,
and more development is needed. However, a successful example is illustrated by the anatomically based model of human
heart ventricles (44). In the following sections, we aim to
provide an overview of some of the methodologies that can be
used to infer biomolecular networks, as well as introduce one
particular approach we have found useful in our research.
•
Synthesis Review
Multilevel Network Integration to Understand Physiology
Topological Analysis of Inferred Biological Networks
Provides Useful Biological Insight
Up to now, we have described some of the most widely used
methodologies for inferring regulatory networks. However, an
immediate challenge arises in interpreting these often large,
complex networks that visually present as a “hairball” (i.e., too
dense a collection of connections to comprehend as a whole)
(40). A simple solution, although not very objective, is to focus
the analysis around a favorite gene(s). In this scenario, the
investigator typically examines the manually selected subnetwork to identify unknown or unexpected biological relationships, which in turn may be used to formulate new hypotheses.
Such “discovery-led” science may be useful when there is
insufficient information to generate hypothesis.
Davidsen PK et al.
301
Alternatively, the topological properties of the network can
be used to identify interesting genes and subnetworks that can
be interpreted. We and others have demonstrated the existence
of a higher-level, modular organization in biological networks
(47, 52, 54), i.e., components of biological systems that act in
collaboration to carry out specific biological processes. Consequently, several modularization approaches have now been
developed to help group subsets of cellular components based
on a given property, such as topological structure or functional
role. Such decomposition of a large complex network into
relatively independent subnetworks (or “modules”) has been
shown to be an effective way to deduce the underlying structure of the fully connected network containing many hundred
variables (so-called “nodes”), as each module can then be
analyzed independently. In addition, studies have demonstrated
that such identified network modules can serve as better predictors of a physiological response than the classic biomarker
discovery approach (see Fig. 1).
In biomolecular interaction networks, as well as subnetworks, nodes have different levels of connectivity (i.e., number
of interactions with other nodes). It has been shown that such
interaction networks have so-called “scale-free” structure properties, as their node connectivity distribution fits a power law
(4). Such a power law degree distribution implies that most of
the connections between biomolecules is linked to a small
number of highly connected nodes, such that a large proportion
of the molecular state of a cell can be explained by a small
subset of biomolecules (so-called “hub” nodes; e.g., a transcription factor that regulates many more genes than average).
Hence, in biological networks a hub is often assumed to be a
key component of a regulatory networks, hence important for
the function of a cell/tissue under investigation. This assumption is supported by the fact that random node disruption does
not significantly affect the network architecture, whereas deletion of hub nodes leads to a complete breakdown of the
network structure (1). Hence, adjusting the spatial position of
each node according to its interconnectivity has been shown to
be a simple, yet effective way of visualizing large complex
interaction networks (57).
More advanced methods to extract information from complex networks exist that aim to identify functional modules
(i.e., subnetworks of biomolecules that are linked to the same
biological function), e.g., by integrating both physical interactions (i.e., experimentally validated protein-protein interactions) and mRNA expression data (27). In this context, an
identified functional module represents a putative multiprotein
complex that is transcriptionally regulated in a specific experimental condition (e.g., treatment vs. control). Hence, by considering additional data on a different level of organization,
one can potentially infer a clearer composite picture of the
underlying biological function.
Finally, to generate objective hypotheses about biological
processes controlled by a specific hub node or subnetwork,
functional enrichment analysis can be performed on all of its
direct neighbors (i.e., all of the adjacent nodes that are directly
connected to the hub) (25). Such enrichment analysis aims at
reducing complexity by defining groups of molecules (represented by gene sets) that share similar biological functions
(e.g., a class of adhesion molecules). To accommodate the
latest advances in knowledge, the different annotation databases used for this purpose [e.g., gene ontology (GO) (2) and
J Appl Physiol • doi:10.1152/japplphysiol.01110.2014 • www.jappl.org
Downloaded from http://jap.physiology.org/ by 10.220.33.1 on June 15, 2017
the distribution of MI values between random permutations of
the original dataset, thereby allowing the elimination of all
nonstatistically relevant dependencies by thresholding using an
appropriate (user-defined) cutoff level. Importantly, the quality
of the inferred interaction network depends on the arbitrarily
selected probability cutoff. A small threshold (e.g., P ⫽ 0.05)
gives a high recall (i.e., fraction of true dependencies that could
be inferred) but low precision, whereas a higher threshold (e.g.,
P ⫽ 10⫺6) yields better precision (i.e., fraction of inferred
dependencies that really are in the network), while suffering
from a low recall. A further advantage of MI as an informationtheoretical measure of dependency between variables concerns
its relatively low computational requirements for building an
interaction network. Hence, MI is able to handle very large
data matrixes with thousands of experimental variables,
whereas most of the other more advanced techniques mentioned (e.g., Bayesian methods) can only deal with much
smaller numbers of variables (⬍100) because of the high
computational complexity. However, to infer robust statistical
associations based on MI, a fairly large sample size is required
(⬎50-100 biological replicates), due to the required estimation
of the (joint) frequency distribution of the connectivity. Interaction networks derived from such reverse engineering methodologies can be visualized and further analyzed using various
freeware software tools, such as Cytoscape (55), Pajak (6), and
BioLayout (18). A comprehensive list of visualization tools
focused on interaction networks and their web-links has recently been reviewed (17).
Up to now, these information-theoretic approaches have
usually been employed on gene expression data only, due to
the wealth of such data available. However, as physiologists
have known for many decades, biological systems are usually
more complex and multilayered. Indeed, despite some popularist science writing to the contrary, genes on their own are
merely permissive elements within biological systems (43).
Furthermore, it has been shown that, when multiple types of
data (e.g., copy number variants, protein, or microRNA expression levels) are incorporated in the network inference
pipeline, the accuracy of the learned network topology increases (49). Hence, at present there is a call for methodologies
that can embed multiple data sources in a single computational
framework. Our recent work has focused on methods that are
able to handle large-scale, multidimensional genomic datasets
(9, 21).
•
Synthesis Review
302
Multilevel Network Integration to Understand Physiology
KEGG (Kyoto Encyclopedia of Genes and Genomes) (45)] are
frequently updated by curators. Using software tools like the
web-based application DAVID (Database for Annotation, Visualization, and Integrated Discovery) (11) or applications
such as BiNGO (36) developed specifically for use with software visualization tools like Cytoscape, one can quickly determine whether any gene sets are statistically over-represented, thus generating hypotheses on the biological processes
controlled by those factors outlined above.
CASE STUDY: INFERENCE OF OXYGEN-DEPENDENT
PATHWAYS IN SKELETAL MUSCLE
Table 1. Anthropometric characteristics defining the COPD
cohort used in the case study
Sex (M/F)
Age, yr
BMI, kg/m2
FFMI, kg/m2
V̇E, l/min
FEV1, liters
FEV1/FVC, %
RV, %predicted
V̇O2max,
l·min⫺1·kg⫺1
Peak power, W
6MWD, m
BODE index
Healthy Controls
COPD, Normal BMI
COPD, Low BMI
10/2
65.3 ⫾ 2.9
26.3 ⫾ 1.1
21.0 ⫾ 0.8
71.2 ⫾ 5.6
3.46 ⫾ 0.2
75.9 ⫾ 2.4
103.9 ⫾ 5.2
9/0
69.4 ⫾ 1.5
27.4 ⫾ 1.4
21.5 ⫾ 0.7
40.5 ⫾ 3.6c
1.41 ⫾ 0.09c
44.0 ⫾ 2.7c
145.0 ⫾ 13.3
6/0
69.2 ⫾ 4.6
19.7 ⫾ 1.0b,e
16.7 ⫾ 0.9b,e
33.0 ⫾ 3.8c
1.21 ⫾ 0.21c
39.5 ⫾ 4.5c
160.0 ⫾ 28.6a
22.3 ⫾ 1.4
117 ⫾ 8
584 ⫾ 24
0.1 ⫾ 0.1
13.9 ⫾ 1.7b
60 ⫾ 7c
469 ⫾ 30a
2.3 ⫾ 0.4b
14.4 ⫾ 1.5b
47 ⫾ 9c
367 ⫾ 59c
4.0 ⫾ 1.0c,d
Values are means ⫾ SE. COPD, chronic obstructive pulmonary disease;
BMI, body mass index; M, male; F, female; FFMI, fat-free mass index; V̇E,
lung ventilation; FEV1, forced expiratory volume in 1 s; FVC, forced vital
capacity; RV, residual volume; V̇O2max, maximum O2 uptake; 6MWD,
6-min walking distance; BODE: BMI, airflow obstruction, dyspnea, and
exercise.aP ⬍ 0.05, bP ⬍ 0.01, and cP ⬍ 0.001 vs. controls. dP ⬍ 0.05 and
e
P ⬍ 0.01 vs. COPD patients with a normal BMI. Comparisons were analyzed
using one-way ANOVA and Tukey’s post hoc test.
Davidsen PK et al.
different computational approach to develop a hierarchical
dynamic model explaining the transcriptional response of oxidative leg muscles to a prolonged gradual reduction in blood
oxygenation (hypoxemia) (Fig. 2, F and G). The model we
describe below validates the notion that the signature identified
using the clinical study may be truly triggered by changes in
oxygen availability. Moreover, the model contributes to the
understanding of the transient events following oxygen depletion that cannot be observed using a cross-sectional clinical
study.
Step 1. Linking Physiological Measurements and Gene
Expression Data in the COPD Cohort
To reconstruct an interaction network spanning multiple
levels of organization, we have utilized the following strategy
that was developed earlier (61).
Combining measurements from different data sources. To
combine gene expression data with whole body physiological
readouts, all variables need to have the same units of measurement (as the range of, e.g., VEGF mRNA expression values are
very different from those of V̇O2max). All such raw scale units
can be unified by simply “transforming” each experimental
variable to have the same dynamic range, e.g., this can be
achieved by standardizing measurements across samples to
have a mean of zero with a SD of 1. Such an established
approach, called z-scoring, enables us to treat the physiological
indicators as individual “nodes” in the inferred interaction
network with states (just as each gene on the array is treated).
DEFINITION OF A BIOLOGICAL FRAMEWORK FOR DATA-DRIVEN
NETWORK INFERENCE. The outcome of data-driven reverse en-
gineering of biological networks, in the absence of any biological assumption(s), often provides results that are difficult to
interpret due to the large number of inferred significant interactions. Thus, to reduce complexity of the problem, we decided to focus the analysis on the set of physiological
parameters and genes encoding for enzymes in the central
bioenergetic pathways (i.e., TCA cycle, oxidative phosphorylation, glycolysis) (see http://pcwww.liv.ac.uk/⬃herberjm/
JAPreview/Table_S2.pdf for additional tables). The latter
choice is reasonable considering the paramount importance of
these molecular pathways in skeletal muscle adaptation. The
overall strategy is, therefore, to identify biomolecules that are
highly correlated (based on MI) with biologically important
experimental variables. Such a focused analysis will generate
multiple network modules of interacting biomolecules, each
with a bioenergetic hub gene or physiological measurement at
its center. Two modules will be linked together if a specific
gene is statistically linked to both hubs.
Reverse engineering. To infer robust regulatory relationships between variables in the integrated multilevel dataset, we
used the ARACNE algorithm. This choice was based on the
large number of measured variables to be considered by the
mathematical framework. By combining all genes expressed in
human skeletal muscle (⬎10,000 mRNAs) with the list of
physiological variables, we far exceed the number of variables
that can be handled by more advanced network inference
methods (e.g., Bayesian methods). Hence, we infer a static
network without any obvious hierarchical organization. The
result of an ARACNE run is an “adjacency matrix” containing
J Appl Physiol • doi:10.1152/japplphysiol.01110.2014 • www.jappl.org
Downloaded from http://jap.physiology.org/ by 10.220.33.1 on June 15, 2017
The main purpose of this case study is to illustrate in a
step-by-step manner the application of reverse engineering to
integrate supracellular physiological measures and genomewide expression profiling. From a more biological perspective,
we aim to identify a clinically relevant signature of hypoxia in
skeletal muscles.
This analysis uses two different datasets. The first is a
publicly available dataset (GSE27536) representing a cohort of
chronic obstructive pulmonary disease (COPD) patients and
healthy controls matched for age and smoking history (10) (see
Table 1 for subject characteristics), which includes gene expression profiling in vastus lateralis muscle and whole body
physiological variables [e.g., V̇O2max, minute ventilation, arterial oxygen tension (PaO2)] (50, 61). The second dataset represents an unpublished, genomewide transcriptional response
of mouse soleus muscle to a gradual decline in atmospheric
oxygen concentration (GSE64076).
Using the first dataset, representing the transcriptional state
of skeletal muscles in a COPD cohort (Fig. 2A), we first show
how to infer connections between oxygen availability (e.g.,
V̇O2max), oxidative stress (protein carbonylation), and gene
expression signatures (Fig. 2, A–C).
Having defined an oxygen-related signature in the disease
setting, we then transpose these findings in a mouse model of
gradual hypoxia (second dataset, Fig. 2, D–E). Here we use a
•
Synthesis Review
Multilevel Network Integration to Understand Physiology
•
Davidsen PK et al.
303
Transcriptomics
COPD patients
with muscle
wasting
Vastus lateralis
sampled at baseline
A
Whole-body
physiological readouts (e.g.
PO2, VO2max,, protein
carbonylation)
B
Static network
inference using
ARACNE
C
D
O2 level
21%
Transcriptomics
E
10%
0
7
F
14
Hypoxic exposure (days)
Hierarchical temporal
model confirms that
oxygen sits on the
highest level of hierarchy
G
Development of dynamical
state-space model
MI values for all pairwise interactions above the specified MI
threshold, which can be visualized automatically in Cytoscape.
After calculating MI-based dependencies between all of the
different variables in our multilevel data matrix, all of those
inferred regulatory interactions with an MI value ⬍ 0.22
(corresponding to a P value cutoff of 10⫺6) were removed.
Such filtering of weaker statistical dependencies is an important step in the generation of a more sparse interaction network,
which can more easily be interpreted by the investigator. The
stringent P value cutoff means that the remaining associations
have been inferred with high precision at the cost of a lower
recall rate.
Network visualization. Data visualized as a network are
often easier to interpret than long lists of biomolecules and
their associated statistical dependencies. Hence, the numeric
output of ARACNE, which contains MI values for all pairwise
associations, was imported into Cytoscape for visualization, a
conventional way of analyzing interaction networks. Briefly,
we reconstruct the network neighborhood of each of the bio-
energetic “seed” genes (i.e., all variables directly connected
to them) (see http://pcwww.liv.ac.uk/⬃herberjm/JAPreview/
Table_S2.pdf for additional tables). The neighboring variables
can either be genes expressed in muscle and/or physiological
variables. Figure 3 summaries key regulatory associations
(based on MI) between this seed set of genes and their
immediate neighbors.
Functional analysis of the network hubs. We further explored whether the direct interacting neighbors of each central
metabolism pathway mapped to functional categories (i.e., GO
terms) as well as KEGG pathways. Notably, a marked enrichment of the different bioenergetic compartments was observed
(Fig. 3, A–C boxes) that clearly highlights the interconnected
nature of the bioenergetic machinery, i.e., functionally related
genes appear to be coexpressed.
Biological interpretation. The most important finding of the
current analysis is that, among the direct neighbors to each
bioenergetic pathway, particularly the two oxidative ones, we
noted a statistical over-representation of genes encoding his-
J Appl Physiol • doi:10.1152/japplphysiol.01110.2014 • www.jappl.org
Downloaded from http://jap.physiology.org/ by 10.220.33.1 on June 15, 2017
Network analysis identifies a
statistical linkage between
oxygen utilization,
inflammation and bioenergetics
Fig. 2. Schematic representation of the analysis strategy used in the case study, highlighting how the
inferred static multiscale network from the clinical
chronic obstructive pulmonary disease (COPD) cohort (A–C) can be bridged to the inference of a
dynamic network representing the temporal progression of events following an experimental challenge
(hypoxic exposure) in a murine animal model (D–G).
Having identified a clinical condition with known
outcome (exercise intolerance in patients with respiratory disease), we could target unknown mechanisms by focusing on one likely source of functional
limitation (skeletal muscle dysfunction ⫾ central
limitation on O2 supply) and generate data characterizing the phenotype. Both genomic and physiological
readouts were used to construct a network of inferred
interactions, which was then interrogated to identify
statistically robust linkages among broad biological
functions. While very useful in providing a list of
useful biomarkers, there remains a potential limitation with single-point associations. The dynamic nature of relationships is captured by repeated measures
across a suitable time scale (which will vary for
different molecular, physiological, and structural
responses) using an animal model of respiratory distress, where the transcriptome-based model demonstrated the central importance of oxygen in the response. V̇O2max, maximum O2 uptake; ARACNE,
Algorithm for the Reconstruction of Accurate Cellular Networks.
Synthesis Review
304
Multilevel Network Integration to Understand Physiology
•
Davidsen PK et al.
A
B
Fig. 3. Graphical representation highlighting putative regulatory associations (significant correlation between two factors is shown as a dashed line) that likely
represent robust interactions, based on high mutual information values. The focus is on central metabolism pathways [i.e., glycolysis, TCA, and oxidative
phosphorylation (OXPHOS), respectively] and their immediate neighbors. The gray boxes define functional enrichment of the different bioenergetic
compartments based on direct neighbors. Individual genes of relevance are grouped into modules with others of related function, as are physiological readouts
that may be treated in a similar manner for statistical analysis. C1–C5: the different complexes in the electron transport chain. The value of such an approach
is in providing a detailed overview of a complex interaction network, reducing the huge number of potential factors into groups of defined function, and offering
a limited number of candidates whose utility as biomarkers or therapeutic targets may be experimentally verified. KEGG, Kyoto Encyclopedia of Genes and
Genomes; GOBP, Gene Ontology Biological Process; GPBP, Goodpasture antigen-binding protein; GOMF, Gene Ontology Molecular Function; PaO2, arterial
oxygen tension; SIRT, sirtuin; HDAC, histone deacetylase.
tone deacetylase (HDAC) enzymes [i.e., HDAC and sirtuin
(SIRT) mRNAs]. This observation is consistent with previous
studies that have highlighted the importance of SIRTs in
regulating metabolism (15, 22, 28). Furthermore, the protein
deacetylase SIRT3 that primarily is localized in the mitochondrial matrix was also significantly positively correlated to both
PaO2 and V̇O2max. In support of deacetylation being an important control point, it was recently shown that Sirt3 knockout
mice exhibit decreased oxygen consumption, thus affecting
cellular respiration (28). Hence, besides the obvious oxygendriven effect on aerobic pathways (as indirect measures of
oxygen availability such as V̇O2max are linked to key genes in
oxidative phosphorylation), the present network-based systems
biology approach points to tissue hypoxia as being a potentially
important player in modifying expression of deacetylase modifying enzymes in severe COPD patients with a muscle-wasting phenotype. Our systems biology approach also negatively
links protein carbonylation [an established proxy measure for
oxidative stress (58)] to complexes 1 and 3 in the electron
transport chain (Fig. 3, bottom left). The validity of such an
association is further strengthened via functional enrichment
analysis using DAVID, as a significant fraction of direct
neighboring genes to protein carbonylation is statistically associated to GO terms representing cellular respiration.
If we then focus on the genes in the glycolytic pathway (Fig.
3, top right), a high proportion of proinflammatory mediators/
receptors (e.g., IL1B, IL1R1, and TNFRSF21) are among the
direct neighbors, as indicated by the enrichment of the “inflammatory response” GO term (Fig. 3A box). Hence, hypoxia
is proinflammatory, as seen by more traditional observation
methods (20).
Multiscale network inference approaches, similar to that
illustrated in Fig. 3, have proven very effective in generating
robust hypotheses (e.g., Ref. 45). However, statistical associations may not represent causality, particularly when the inferred associations stem from steady-state measures. Thus, to
validate our hypothesis that varying oxygen levels (represented
by V̇O2max and PaO2) control the expression of epigenetic
J Appl Physiol • doi:10.1152/japplphysiol.01110.2014 • www.jappl.org
Downloaded from http://jap.physiology.org/ by 10.220.33.1 on June 15, 2017
C
Synthesis Review
Multilevel Network Integration to Understand Physiology
modifiers, we used a more sophisticated network inference
algorithm that can learn the structure of networks from timecourse data. We applied this dynamic inference approach to a
murine model of hypoxia (step 2).
O2 level
30
20
10
PC2 (15%)
21%
10%
0
-10
-20
-30
0
7
14
Hypoxic exposure (days)
-40
-20
0
20
40
PC1 (53%)
C
Fig. 4. High-level representation of temporal transcriptional changes in the murine model of hypoxia. A: graphical representation of the preclinical experimental
design. The oxygen level was gradually decreased from 21% to 10% during the first week, and mice were housed for another week at this oxygen concentration.
B: principal component (PC) plot highlighting the transcriptional dynamics caused by the hypoxic challenge. C: hierarchical clustering using mRNA expression
levels of genes modulated by hypoxia (P ⬍ 0.05). Each row represents a transcript, and each column represents a sample. Red and green colors indicate
expression levels above and below the median value of the distribution of signal, respectively. Using solid yellow lines, we have subdivided genes into overall
trends to help the reader. Enriched functional terms within these are listed next to the heatmap. GOCC, Gene Ontology Cellular Component.
J Appl Physiol • doi:10.1152/japplphysiol.01110.2014 • www.jappl.org
Downloaded from http://jap.physiology.org/ by 10.220.33.1 on June 15, 2017
Animal models are commonly used for studying the in vivo
effects of hypoxia, for ethical reasons, where severe or prolonged hypoxemia is induced and invasive samples are required to explore mechanisms. Importantly, hindlimb skeletal
muscles have been reported to alter metabolic phenotype and
reduce fiber size in response to prolonged hypoxic stress in
mice (53, 63), highlighting their potential relevance as a
preclinical model of muscle wasting in COPD patients. To
experimentally test the hypothesis derived from the clinical
COPD network presented in Fig. 3, we, therefore, exposed
adult male C57/Bl6 mice to chronic systemic hypoxia for up to
2 wk, to simulate levels of hypoxemia reported in COPD
patients with advanced respiratory insufficiency. To capture
B
305
Davidsen PK et al.
the temporal effect of reduced oxygen tension on gene regulation, we sampled and gene profiled the soleus muscle (n ⫽ 4)
at three different time points (days 3, 7, and 14) following
initiation of the gradual hypoxic insult (i.e., the O2 level was
gradually lowered to 10% over the first week and kept stable
during the second week) (Fig. 2, bottom).
First, a high-level representation of the temporal transcriptional changes was performed using a variable reduction technique called principal component analysis (PCA) (Fig. 4B).
When plotting replicates of two variables against each other, it
is relatively easy to see which is a better discriminating factor.
Visual inspection becomes increasingly difficult as the number
of variables increase, hence the need for PCA. In essence, this
method aims at “tilting” the axes through the multidimensional
data space, such that the first principal component accounts for
as much of the variation in the original dataset as possible (the
assumption is that the most important dynamics in the dataset
are the ones with the largest variation). Our PCA revealed that
the early dynamics of hypoxia are captured by the first principal component, whereas the second most important principal
Step 2. Gene Expression Dynamics in Response to Tissue
Hypoxia
A
•
Synthesis Review
306
Multilevel Network Integration to Understand Physiology
regulatory interactions (e.g., levels of regulatory proteins as
well as effects of mRNA and protein degradation).
The next step was to apply state-space modeling to reverse
engineer transcriptional network modules (i.e., representing
discrete temporal dimensions) from our replicated murine time
course dataset. Such module-based reduction in complexity
allows analysis of hundreds or even thousands of genes, as
those with a similar temporal expression profile are aggregated
into a transcriptional module. To allow construction of a near
genome-level model, we took advantage of a newer approach
that incorporates this concept of modularization (23).
A SSM can reconstruct the topology of a network representing
the systems dynamics, despite a relatively small number of time
points, by using biological replicates for each time point (23). To
reduce complexity, variables that do not change significantly are
excluded from the modeling process. In this case study, genes
deemed to be significant by ANOVA at a 1% significance level,
as well as all hub genes, were included (931 variables in total) (see
http://pcwww.liv.ac.uk/⬃herberjm/JAPreview/Table_S3.pdf for
+
1.5
1.0
0.5
Transcription
0.0
-0.5
+
-1.0
1.5
-1.5
1.0
-
0.5
1.5
Endocytosis
Immune response
Oxygen
1.0
0.0
0.5
-0.5
0.66
0.0
-0.5
-
-0.43
-1.5
-0.63
-1.0
-1.5
-1.0
-0.38
Inflammation
HIF-1alpha
1.5
1.0
Module 1
Muscle contraction
OXPHOS
Glycolysis
SIRT3 + SIRT5
0.5
0.0
-0.5
+
1.5
-1.0
1.0
-1.5
0.5
Inflammation
0.0
Module 2
0.44
-0.5
-0.33
0.66
-1.0
-1.5
-
mTOR signal.
Insulin signal.
OXPHOS
SIRT6
+
1.5
1.5
1.0
1.0
0.5
0.5
0.0
0.0
-0.5
-0.5
Muscle differentiation
HDAC4, -5 & -9
-1.0
-1.0
-1.5
-1.5
Module 3
-
1.5
1.0
Blood vessel development
Tissue remodeling
Neg. regulation of differentiation
0.5
0.0
-0.5
-1.0
-1.5
Module 4
Fig. 5. The hierarchical dynamic state-space model identified four modules (x-axes define length of hypoxic exposure), each characterized by two separate
transcriptional profiles: plus and minus, representing up- and downregulation, respectively. The hierarchical position of the modules represents the estimated
temporal structure of the network. Functionally enriched gene ontology (GO) terms (regular text) as well as key genes (italics) are identified next to the relevant
module. Blue arrows represent temporal repression, whereas red arrows represent temporal induction. The numerical value next to each arrow represents the
estimated coefficient. mTOR, mammalian target of rapamycin; HIF-1␣, hypoxia-inducible factor-1␣.
J Appl Physiol • doi:10.1152/japplphysiol.01110.2014 • www.jappl.org
Downloaded from http://jap.physiology.org/ by 10.220.33.1 on June 15, 2017
component (in terms of variance captured) separated the later
time points. Furthermore, functional enrichment analysis of the
differentially expressed genes (ANOVA, P ⬍ 0.05) using
DAVID (Fig. 4A) highlighted several important pathways/
ontologies. Most striking was the enrichment of protein catabolic process and ubiquitin-mediated proteolysis among genes
upregulated at days 7 and 14, clearly suggestive of a transcriptionally regulated muscle wasting phenotype driven by the
experimentally induced hypoxemic state.
State space models (SSMs) are a class of probabilistic
graphical models (33). SSM provides a general framework for
analyzing deterministic and stochastic dynamic systems that
can be measured/observed through a stochastic process. The
SSM framework has been successfully used for the analysis of
gene expression data (23, 51). In its simpler application, the
model formalizes the effect of hidden, unmeasurable factors in
specifying observed gene expression changes over time. The
inclusion of these hidden factors is important, since we cannot
hope to measure all possible factors contributing to genetic
Davidsen PK et al.
•
Synthesis Review
Multilevel Network Integration to Understand Physiology
Davidsen PK et al.
307
tion, tissue remodeling, and blood vessel development. Interestingly, three HDACs are represented in module 4(⫹) (Fig. 5).
Figure 6 represents a more focused version of Fig. 5, highlighting the most significant interactions between components
in the four inferred modules from Fig. 5.
We, therefore, conclude that the inferred dynamic model
using a SSM approach appropriately recapitulates the interpretative model advanced in Fig. 3. In addition, it identifies
oxygen at the highest level of hierarchy, whereas key
effector functions controlled by oxygen, such as inflammation and muscle differentiation, are downstream in the
temporal hierarchy.
CONCLUSIONS
The aim of this brief review is to provide an intuitive
overview on data-driven “learning” of biological pathways,
linking molecular and physiological readouts. We used a case
study to make it easier for experimental biologists to see the
potential of computational biology to provide interpretative
models of complex patterns, and stress that the identification of
general properties of a system from a genomewide analysis of
a molecular state of a system is a very powerful approach.
The ability to generate omics data with relatively accessible
technologies offers an unprecedented opportunity to study how
Fig. 6. A higher resolution representation of Fig.
5, highlighting the most significant gene interactions between components in the four inferred
modules, is shown. Lines represent factor interactions based on mutual information (blue represents temporal repression, red represents temporal induction). Genes are color coded for
broad functional categories (red, cytokines;
blue, epigenetic modifiers; green, aerobic metabolism; purple, muscle differentiation; yellow,
cell interaction).
J Appl Physiol • doi:10.1152/japplphysiol.01110.2014 • www.jappl.org
Downloaded from http://jap.physiology.org/ by 10.220.33.1 on June 15, 2017
additional tables). The hub genes were chosen to represent the
different components in our interpretative model derived from the
clinical COPD dataset (Fig. 3). Finally, the experimentally set
oxygen level was used as an independent variable.
Based on unsupervised clustering using HOPACH within
the software programming environment R (65), we identified
eight distinct gene clusters with similar expression profiles.
Hence, to model the effect of hypoxemia on the skeletal muscle
transcriptome, the hidden state dimension was set to 4, as each
inferred module contains both a positive (⫹) and a negative
(⫺) component.
The hierarchical dynamic model in four temporal dimensions shows that modules 1 and 2, which sit on the highest level
of hierarchy (i.e., precede others in time), were enriched in GO
terms related to muscle contraction, bioenergetic pathways,
and inflammation, among others (Fig. 5). Interestingly, the
experimental oxygen concentration was represented in module
1(⫺), whereas two deacetylases SIRT3 and SIRT5 were found
in module 2(⫺). A negative influence is observed of module 1
on module 3, which is located further down the temporal
hierarchy. Module 3(⫹) is highly enriched in inflammatory
processes, whereas its negative counterpart mainly represents
two key signaling pathways (mammalian target of rapamycin
and insulin). At the lowest temporal level we find module 4,
which is enriched in GO terms related to muscle differentia-
•
Synthesis Review
308
Multilevel Network Integration to Understand Physiology
ACKNOWLEDGMENTS
At the request of the author(s), readers are herein alerted to the fact that
additional materials related to this manuscript may be found at the institutional
website of one of the authors, which is: http://pcwww.liv.ac.uk/⬃herberjm/
Davidsen PK et al.
JAPreview/Table_S2.pdf and http://pcwww.liv.ac.uk/⬃herberjm/JAPreview/
Table_S3.pdf. These materials are not a part of this manuscript and have not
undergone peer review by the American Physiological Society (APS). APS and
the journal editors take no responsibility for these materials, for the website
address, or for any links to or from it.
DISCLOSURES
No conflicts of interest, financial or otherwise, are declared by the author(s).
AUTHOR CONTRIBUTIONS
Author contributions: P.K.D. and F.F. conception and design of research;
P.K.D., S.E., and F.F. performed experiments; P.K.D. and F.F. analyzed data;
P.K.D. and F.F. interpreted results of experiments; P.K.D., N.T., and F.F.
prepared figures; P.K.D., S.E., and F.F. drafted manuscript; P.K.D., S.E., and
F.F. edited and revised manuscript; P.K.D., S.E., and F.F. approved final
version of manuscript.
REFERENCES
1. Alderson D, Doyle JC, Li L, Willinger W. Towards a theory of
scale-free graphs: definition, properties, and implications. Internet Math 2:
431–523, 2005.
2. Ashburner M, Ball CA, Blake JA, Botstein D, Butler H, Cherry JM,
Davis AP, Dolinski K, Dwight SS, Eppig JT, Harris MA, Hill DP,
Issel-Tarver L, Kasarskis A, Lewis S, Matese JC, Richardson JE, Ringwald M, Rubin GM, Sherlock G. Gene ontology: tool for the unification of
biology. The Gene Ontology Consortium. Nat Genet 25: 25–29, 2000.
3. Bansal M, Della Gatta G, di Bernardo D. Inference of gene regulatory
networks and compound mode of action from time course gene expression
profiles. Bioinformatics 22: 815–822, 2006.
4. Barabasi A, Albert R. Emergence of scaling in random networks. Science
286: 509 –512, 1999.
5. Barrett T, Suzek TO, Troup DB, Wilhite SE, Ngau WC, Ledoux P,
Rudnev D, Lash AE, Fujibuchi W, Edgar R. NCBI GEO: mining
millions of expression profiles– database and tools. Nucleic Acids Res 33:
D562–D566, 2005.
6. Batagelj V, Mrvar A. Pajek–program for large network analysis. Connect
(Tor) 21: 47–57, 1998.
7. Brazma A, Parkinson H, Sarkans U, Shojatalab M, Vilo J, Abeygunawardena N, Holloway E, Kapushesky M, Kemmeren P, Lara GG,
Oezcimen A, Rocca-Serra P, Sansone SA. ArrayExpress–a public repository for microarray gene expression data at the EBI. Nucleic Acids Res
31: 68 –71, 2003.
8. Butte AJ, Kohane IS. Mutual information relevance networks: functional
genomic clustering using pairwise entropy measurements. Pac Symp
Biocomput 2000: 418 –429, 2000.
9. Cassese A, Guindani M, Tadesse MG, Falciani F, Vannucci M. A
hierarchical Bayesian model for inference of copy number variants and
their association to gene expression. Ann Appl Stat 8: 148 –175, 2014.
10. Davidsen PK, Herbert JM, Antczak P, Clarke K, Ferrer E, Peinado
VI, Gonzalez C, Roca J, Egginton S, Falciani F. A systems biology
approach reveals a link between systemic cytokines and skeletal muscle
energy metabolism in a rodent smoking model and human COPD. Genome
Med 6: 59, 2014.
10a.De Smet R, Marchal K. Advantages and limitations of current network
inference methods. Nat Rev Microbiol 8: 717–729, 2010.
11. Dennis G, Sherman BT, Hosack DA, Yang J, Gao W, Lane HC,
Lempicki RA. DAVID: Database for Annotation, Visualization, and
Integrated Discovery. Genome Biol 4: P3, 2003.
12. Du W, Elemento O. Cancer systems biology: embracing complexity to develop
better anticancer therapeutic strategies. Oncogene 34: 3215–3225, 2015.
13. Faith JJ, Hayete B, Thaden JT, Mogno I, Wierzbowski J, Cottarel G,
Kasif S, Collins JJ, Gardner TS. Large-scale mapping and validation of
Escherichia coli transcriptional regulation from a compendium of expression profiles. PLoS Biol 5: e8, 2007.
14. Falati S, Gross P, Merrill-Skoloff G, Furie BC, Furie B. Real-time in
vivo imaging of platelets, tissue factor and fibrin during arterial thrombus
formation in the mouse. Nat Med 8: 1175–1180, 2002.
15. Finkel T, Deng CX, Mostoslavsky R. Recent progress in the biology and
physiology of sirtuins. Nature 460: 587–591, 2009.
16. Gavaghan D, Garny A, Maini PK, Kohl P. Mathematical models in
physiology. Philos Trans A Math Phys Eng Sci 364: 1099 –1106, 2006.
J Appl Physiol • doi:10.1152/japplphysiol.01110.2014 • www.jappl.org
Downloaded from http://jap.physiology.org/ by 10.220.33.1 on June 15, 2017
genetic information is used to control complex biological
processes and their interaction. Until now we have only been
able to understand a fraction of that complexity. The computational methods described in this review are designed to
support this effort in the measure that they help isolate from
these large datasets molecular signatures that correlate to
phenotypic outcome.
With the help of computational biology, we are, therefore,
able to develop hypotheses, which can be experimentally
validated. In this context data-driven biology is not in contraposition with hypothesis-driven research. Instead it is a tool
that supports hypothesis generation in the event that the data
are too complex to be interpreted solely using common sense.
This approach is well developed in other areas of science, such
as cancer biology, where there is a vast literature showing that
important hypotheses can be generated from modeling of these
large datasets (12, 62).
In this paper, we demonstrate the development of an integrative workflow that incorporates measurements from different levels of cellular and molecular organization using a case
study representing muscle wasting in COPD. The outline
provides an exemplar where individual steps can be modified
according to the type of data at hand, and additional data types
added. For example, in contrast to established gene expression
microarrays, techniques for proteomics and especially metabolomics are still under development. Once it is possible to
measure the whole proteome and metabolome of a sample,
systems identification pipelines will clearly benefit from these
omics techniques.
The specific findings in the case study relate to the definition
of an oxygen-dependent signature in COPD. Such signature
(exemplified in Fig. 3) is static and entirely based on statistical
inference. The model is, therefore, based only on correlation
between a series of patient biopsy snapshots and, therefore,
does not allow any inference of causality. The use of a mouse
model of gradual hypoxia allowed us to demonstrate that a
signature inferred from the clinical cohort is indeed modulated
by experimental reduction in oxygen levels. Moreover, the
development of a mathematical model identifies oxygen as the
most upstream event as an emergent property. This may appear
as an obvious finding, but, from a methodological perspective,
validates the analytic approach.
The data we have used in this case study are gene expression
profiling and as such are representative of available datasets.
This has several limitations. The first is that models, including
multiple levels in the expression of genetics information (e.g.,
epigenetics, microRNA, proteomics, metabolomics, etc.) may
better represent biological complexity. However, current computational methods are inadequate to represent properly the
interaction between these levels. Moreover, time course data
that rely on disruptive sampling strategies are not true time
course experiments. As the new functional genomics technologies develop further, as well as novel approaches to model the
interaction between different layer of biological organization,
we expect that the efficacy of data-driven approaches will
increase further.
•
Synthesis Review
Multilevel Network Integration to Understand Physiology
Davidsen PK et al.
309
40. Merico D, Gfeller D, Bader GD. How to visually interpret biological data
using networks. Nat Biotechnol 27: 921–924, 2009.
41. Meyer PE, Kontos K, Lafitte F, Bontempi G. Information-theoretic
inference of large transcriptional regulatory networks. EURASIP J Bioinform Syst Biol 2007: 79879, 2007.
42. Neapolitan RE. Learning Bayesian Networks. Upper Saddle River, NJ:
Pearson Prentice Hall, 2004.
43. Noble D. The Music of Life: Biology Beyond Genes. Oxford, UK: Oxford
University Press, 2008.
44. Noble D. Computational models of the heart and their use in assessing the
actions of drugs. J Pharm Sci 107: 107–117, 2008.
45. Ogata H, Goto S, Sato K, Fujibuchi W, Bono H, Kanehisa M. KEGG:
Kyoto Encyclopedia of Genes and Genomes. Nucleic Acids Res 27:
29 –34, 1999.
46. Opgen-Rhein R, Strimmer K. Learning causal networks from systems
biology time course data: an effective model selection procedure for the
vector autoregressive process. BMC Bioinformatics 8, Suppl 2: S3, 2007.
47. Ortega F, Sameith K, Turan N, Compton R, Trevino V, Vannucci M,
Falciani F. Models and computational strategies linking physiological
response to molecular networks from large-scale data. Philos Trans A
Math Phys Eng Sci 366: 3067–3089, 2008.
48. Perrin BE, Ralaivola L, Mazurie A, Bottani S, Mallet J, d’Alché-Buc
F. Gene networks inference using dynamic Bayesian networks. Bioinformatics 19, Suppl 2: ii138 –ii148, 2003.
49. Poultney CS, Greenfield A, Bonneau R. Integrated inference and analysis of regulatory networks from multi-level measurements. Methods Cell
Biol 110: 19 –56, 2012.
50. Rabinovich RA, Bastos R, Ardite E, Llinàs L, Orozco-Levi M, Gea J,
Vilaró J, Barberà JA, Rodríguez-Roisin R, Fernández-Checa JC,
Roca J. Mitochondrial dysfunction in COPD patients with low body mass
index. Eur Respir J 29: 643–650, 2007.
51. Rangel C, Angus J, Ghahramani Z, Lioumi M, Sotheran E, Gaiba A,
Wild DL, Falciani F. Modeling T-cell activation using gene expression
profiling and state-space models. Bioinformatics 20: 1361–1372, 2004.
52. Ravasz E, Somera AL, Mongru DA, Oltvai ZN, Barabási AL. Hierarchical organization of modularity in metabolic networks. Science 297:
1551–1555, 2002.
53. Reinke C, Bevans-Fonti S, Drager LF, Shin MK, Polotsky VY. Effects
of different acute hypoxic regimens on tissue oxygen profiles and metabolic outcomes. J Appl Physiol 111: 881–890, 2011.
54. Segal E, Shapira M, Regev A, Pe’er D, Botstein D, Koller D, Friedman N.
Module networks: identifying regulatory modules and their condition-specific
regulators from gene expression data. Nat Genet 34: 166 –176, 2003.
55. Shannon P, Markiel A, Ozier O, Baliga NS, Wang JT, Ramage D, Amin N,
Schwikowski B, Ideker T. Cytoscape: a software environment for integrated
models of biomolecular interaction networks. Genome Res 13: 2498–2504, 2003.
57. Su G, Kuchinsky A, Morris JH, States DJ, Meng F. GLay: community
structure analysis of biological networks. Bioinformatics 26: 3135–3137, 2010.
58. Suzuki YJ, Carini M, Butterfield DA. Protein carbonylation. Antioxid
Redox Signal 12: 323–325, 2010.
59. Timmons JA, Knudsen S, Rankinen T, Koch LG, Sarzynski M, Jensen
T, Keller P, Scheele C, Vollaard NBJ, Nielsen S, Akerström T,
MacDougald OA, Jansson E, Greenhaff PL, Tarnopolsky MA, van
Loon LJC, Pedersen BK, Sundberg CJ, Wahlestedt C, Britton SL,
Bouchard C. Using molecular classification to predict gains in maximal
aerobic capacity following endurance exercise training in humans. J Appl
Physiol 108: 1487–1496, 2010.
60. Trevino V, Falciani F. GALGO: an R package for multivariate variable
selection using genetic algorithms. Bioinformatics 22: 1154 –1156, 2006.
61. Turan N, Kalko S, Stincone A, Clarke K, Sabah A, Howlett K,
Curnow SJ, Rodriguez DA, Cascante M, O’Neill L, Egginton S, Roca
J, Falciani F. A systems biology approach identifies molecular networks
defining skeletal muscle abnormalities in chronic obstructive pulmonary
disease. PLoS Comput Biol 7: e1002129, 2011.
62. Werner HMJ, Mills GB, Ram PT. Cancer systems biology: a peek into
the future of patient care? Nat Rev Clin Oncol 11: 167–176, 2014.
63. Willmann G. Transcriptional Regulation after Chronic Hypoxia Exposure in Skeletal Muscle. Cologne, Germany: University of Cologne, 2013.
64. Yu J, Smith VA, Wang PP, Hartemink AJ, Jarvis ED. Advances to
Bayesian network inference for generating causal networks from observational biological data. Bioinformatics 20: 3594 –3603, 2004.
65. van der Laan MJ, Pollard KS. Hybrid clustering of gene expression data with
visualization and the bootstrap. J Stat Plan Inference 117: 275–303, 2003.
J Appl Physiol • doi:10.1152/japplphysiol.01110.2014 • www.jappl.org
Downloaded from http://jap.physiology.org/ by 10.220.33.1 on June 15, 2017
17. Gehlenborg N, O’Donoghue SI, Baliga NS, Goesmann A, Hibbs MA,
Kitano H, Kohlbacher O, Neuweger H, Schneider R, Tenenbaum D,
Gavin AC. Visualization of omics data for systems biology. Nat Methods
7: S56 –S68, 2010.
18. Goldovsky L, Cases I, Enright AJ, Ouzounis CA. BioLayout (Java):
versatile network visualisation of structural and functional relationships.
Appl Bioinformatics 4: 71–74, 2005.
19. Gomez-Cabrero D, Compte A, Tegner J. Workflow for generating
competing hypothesis from models with parameter uncertainty. Interface
Focus 1: 438 –449, 2011.
20. Gonzalez NC, Wood JG. Alveolar hypoxia-induced systemic inflammation: what low PO(2) does and does not do. Adv Exp Med Biol 662: 27–32,
2010.
21. Gupta R, Stincone A, Antczak P, Durant S, Bicknell R, Bikfalvi A,
Falciani F. A computational framework for gene regulatory network
inference that combines multiple methods and datasets. BMC Syst Biol 5:
52, 2011.
22. He W, Newman JC, Wang MZ, Ho L, Verdin E. Mitochondrial sirtuins:
regulators of protein acylation and metabolism. Trends Endocrinol Metab
23: 467–476, 2012.
23. Hirose O, Yoshida R, Imoto S, Yamaguchi R, Higuchi T, CharnockJones DS, Print C, Miyano S. Statistical inference of transcriptional
module-based gene networks from time course gene expression profiles by
using state space models. Bioinformatics 24: 932–942, 2008.
24. Holloway KV, O’Gorman M, Woods P, Morton JP, Evans L, Cable
NT, Goldspink DF, Burniston JG. Proteomic investigation of changes in
human vastus lateralis muscle in response to interval-exercise training.
Proteomics 9: 5155–5174, 2009.
25. Huang DW, Sherman BT, Lempicki RA. Bioinformatics enrichment
tools: paths toward the comprehensive functional analysis of large gene
lists. Nucleic Acids Res 37: 1–13, 2009.
26. Huxley AF. Muscle structure and theories of contraction. Prog Biophys
Biophys Chem 7: 255–318, 1957.
27. Ideker T, Ozier O, Schwikowski B, Siegel AF. Discovering regulatory
and signalling circuits in molecular interaction networks. Bioinformatics
18, Suppl 1: S233–S240, 2002.
28. Jing E, Emanuelli B, Hirschey MD, Boucher J, Lee KY, Lombard D,
Verdin EM, Kahn CR. Sirtuin-3 (Sirt3) regulates skeletal muscle metabolism and insulin signaling via altered mitochondrial oxidation and
reactive oxygen species production. Proc Natl Acad Sci U S A 108:
14608 –14613, 2011.
29. Jones P, Côté RG, Martens L, Quinn AF, Taylor CF, Derache W,
Hermjakob H, Apweiler R. PRIDE: a public repository of protein and
peptide identifications for the proteomics community. Nucleic Acids Res
34: D659 –D663, 2006.
30. Joyner MJ, Pedersen BK. Ten questions about systems biology. J
Physiol 589: 1017–1030, 2011.
31. Keller P, Vollaard NBJ, Gustafsson T, Gallagher IJ, Sundberg CJ,
Rankinen T, Britton SL, Bouchard C, Koch LG, Timmons JA. A
transcriptional map of the impact of endurance exercise training on
skeletal muscle phenotype. J Appl Physiol 110: 46 –59, 2011.
32. Kohl P, Noble D. Systems biology and the virtual physiological human.
Mol Syst Biol 5: 292, 2009.
33. Koller D, Friedman N. Probabilistic Graphical Models. Cambridge,
MA: MIT Press, 2009.
34. Krogh A. The number and distribution of capillaries in muscles with
calculations of the oxygen pressure head necessary for supplying the
tissue. J Physiol 52: 409 –415, 1919.
35. Kupershmidt I, Su QJ, Grewal A, Sundaresh S, Halperin I, Flynn J,
Shekar M, Wang H, Park J, Cui W, Wall GD, Wisotzkey R, Alag S,
Akhtari S, Ronaghi M. Ontology-based meta-analysis of global collections of high-throughput public data. PLoS One 5: e13066, 2010.
36. Maere S, Heymans K, Kuiper M. BiNGO: a Cytoscape plugin to assess
overrepresentation of gene ontology categories in biological networks.
Bioinformatics 21: 3448 –3449, 2005.
37. Mah N. A comparison of oligonucleotide and cDNA-based microarray
systems. Physiol Genomics 16: 361–370, 2004.
38. Margolin AA, Nemenman I, Basso K, Wiggins C, Stolovitzky G, Dalla
Favera R, Califano A. ARACNE: an algorithm for the reconstruction of
gene regulatory networks in a mammalian cellular context. BMC Bioinformatics 7, Suppl 1: S7, 2006.
39. Mattson DL. Functional genomics. In: Integrative Physiology in the
Proteomics and Post-Genomics Age, edited by Walz W. New York:
Humana, 2005, p. 7–26.
•